My pi setup. Themes, prompts, agent instructions, and a few opinions about how an AI coding agent should behave.
Fork this repo, rip out my opinions, and replace them with yours.
Most pi setups start as a single settings.json and slowly grow into a mess of prompts, themes, and half-configured skills scattered across your home directory. This repo is the fix — version-controlled, portable, and opinionated enough to be useful on day one.
I wanted my agent to:
- Stop asking obvious questions — The
AGENTS.mdgives it enough context to just get to work - Match my workflow — Prompt templates for the things I actually do: debug, review, plan, refactor
- Not look like a default terminal — Three custom themes because staring at a screen all night means it better look good
agent/
├── AGENTS.md # Who I am, how I work, what I expect
├── settings.json # Model, theme, packages, preferences
├── compact-config.json # Context window management
├── extensions/
│ └── double-escape.ts # Double-press Escape to abort (not single)
├── prompts/ # Reusable prompt templates
│ ├── debug.md # "Don't guess. Investigate systematically."
│ ├── explain.md # Break down code/systems clearly
│ ├── plan.md # Plan before you touch code
│ ├── refactor.md # Change structure, preserve behavior
│ └── review.md # Find bugs, not style nitpicks
└── themes/
├── zosma-classic.json # Catppuccin-inspired. The daily driver.
├── zosma-dark.json # Tailwind palette. High contrast.
└── zosma-cyber.json # Neon everything. For those nights.
The AGENTS.md is the most important file here. It tells pi:
- Who it's working with — Stack preferences, org context, toolchain
- How to write code — TypeScript first, small functions, comments only for the "why"
- How to communicate — Be direct, explain trade-offs, flag risks
- How to work — Read before writing, test after changing, investigate before fixing
This isn't a generic "be helpful" prompt. It's a working contract between me and the agent. The more specific you make yours, the less time you spend correcting it.
Fork note: This is the file you should customize first. Replace my context with yours.
Templates for the workflows I use daily. Not fancy, just useful.
| Template | What It Does |
|---|---|
debug |
Systematic debugging — reproduce → hypothesize → verify → fix |
explain |
High-level overview + key components + design decisions |
plan |
Full implementation plan with rollback strategy |
refactor |
Safe restructuring — same behavior, better code |
review |
Bug hunting with file locations and line numbers |
Three themes, three moods.
- zosma-classic — Catppuccin Mocha palette with lavender headings and teal accents. Soft purples, warm peaches. The one I use 90% of the time.
- zosma-dark — Built on Tailwind's color system. Cyan accents on a near-black base. Clean and high contrast.
- zosma-cyber — Neon cyan, neon purple, neon green on a pitch-black background. Looks ridiculous. Feels great at 3am.
Custom extensions in agent/extensions/:
| Extension | What It Does |
|---|---|
double-escape.ts |
Requires pressing Escape twice within 500ms to abort — first press shows a hint in the footer, second press aborts. No more accidental cancellations. |
Community packages that extend pi's capabilities:
| Package | Why |
|---|---|
| pi-superpowers | Brainstorming, TDD, code review, systematic debugging workflows |
| pi-extmgr | Extension manager |
# Clone into your home directory
git clone git@github.com:arjun-zosma/dotpi.git ~/.pi
# Install community packages
pi install
# Add your credentials (never committed)
# Create agent/auth.json with your API keys| Path | Why |
|---|---|
agent/auth.json |
Your API keys. Obviously. |
agent/sessions/ |
Conversation history |
agent/git/ |
Cloned packages (reinstalled via pi install) |
agent/skills/pi-skills/ |
Installed skills (managed separately) |
agent/bin/ |
Installed binaries |
agent/.extmgr-cache/ |
Extension manager cache |
MIT — Take what's useful, make it yours.